Video analytic system on FPGA edge device for real time fire detection

Abstract

This Final Year Project aims to develop a video-based fire detection system on Xilinx Kira KV260 evaluation board with FPGA SoC. State-of-the-art Yolov8 is used as the base architecture to develop fire detection model. To address the limitations of small training dataset, various techniques such as data augmentation, CLAHE image processing, and Squeeze-and-Excitation blocks are examined and then selected to enhance model performance. Knowledge distillation, a model compression technique, is used in the form of self-distillation to further increase detection accuracy. The developed detection model undergoes hardware adaptive adjustment and quantization for the embedded deployment.